🚀 RELEASES & LAUNCHES
- Agility unveils Digit 5 humanoid designed to work safely near people
Agility unveiled Digit 5, a humanoid robot designed for safe work alongside humans.- Target users are warehouse and automotive-factory operators, including GXO, Schaeffler, Amazon, and Toyota Motor Manufacturing Canada.
- Digit 5 detects nearby people and can move away, stop, or squat; early access is planned for the first half of 2027 and general availability by year-end.
- The robot lifts up to 50 pounds, runs for 90 minutes, recharges in nine minutes, and reaches up to 7.2 feet.
- Digit 5 uses Nvidia Thor IGX hardware, Nvidia Halos for Robotics, multimodal vision sensors, and swappable grippers.
- Its cooperative-safety features are designed to let robots work without physical cages or barriers in shared workplace areas.
🔬 RESEARCH
- IBM Research targets AI-agent inconsistency with new evaluation method
IBM Research introduced tools to measure and reduce AI-agent inconsistency.- For builders deploying ReAct agents in production workflows where repeated requests must produce reliable results.
- IBM Research added a Consistency Analyzer and consistency guidelines to ALTK-Evolve.
- On AppWorld, GPT-4.1 achieved 77.4% Mean@5 but only 53.0% Pass^5, a 24.4-point consistency gap.
- The analyzer resamples each decision point from one recorded trajectory with five completions by default and needs no ground truth.
- Consistency guidelines reduced the gap from 24.4 points to 12.0 points without reducing average accuracy.
Sources 📎 Coverage: huggingface.co · 👁 via Hugging Face Blog
📊 BENCHMARKS & EVALS
- Swift-Qwen3.8-27B cuts Qwen reasoning tokens by 58.3%
UkisAI released Swift-Qwen3.8-27B, a reasoning-efficient Qwen3.8 derivative.- For developers running Qwen3.8-27B locally or through Hugging Face-compatible tools.
- The 27B model uses 58.3% fewer thinking tokens with under 1% performance loss.
- It reports up to 1.95× speed-ups on several tasks.
- Swift fine-tunes Qwen by penalizing reasoning-marker tokens and incorporates transfer from ThinkingCap-Qwen3.6-27B.
- The model is available on Hugging Face under the Swift Open License 1.0.
💰 BUSINESS & FUNDING
OpenAI, Anthropic and Google DeepMind confirm weeks of AI safety talks
OpenAI, Anthropic and Google DeepMind have held weeks of AI safety discussions.- The talks involve OpenAI, Anthropic and Google DeepMind, three frontier AI developers.
- OpenAI said the companies have discussed AI safety for several weeks, including industry standards and independent model evaluations.
- The coordination comes amid debate over catastrophic AI risks, antitrust concerns and U.S. regulation.
Sources 📎 Coverage: techcrunch.com · 👁 via TechCrunch AIUS data centers could consume 18B cubic feet of gas daily by 2035
BloombergNEF projects US data centers will consume 18 billion cubic feet of natural gas daily by 2035.- AI companies, data-center operators, utilities, and ratepayers would be affected, with Meta, Microsoft, Google, and Amazon planning onsite gas power.
- BloombergNEF projects US data centers will consume about 18 billion cubic feet of natural gas per day by 2035, nearly double its forecast from nine months earlier.
- Grid-connected data centers are projected to add 15 billion cubic feet per day of gas demand, while onsite facilities would consume 2.9 billion to 3.4 billion.
- The growth could raise natural-gas prices and add about 1 million metric tons of greenhouse-gas pollution daily.
Sources 📎 Coverage: techcrunch.com · 👁 via TechCrunch AI